GLM-5.2 adopts critic-based PPO training

While the industry recently rushed toward group-based reinforcement learning (like GRPO) to save on critic-model compute overhead, Zhipu's pivot back to critic-based PPO shows that group-wise variance reduction fails to scale for ultra-long contexts and complex, multi-turn reasoning.

• –The Critic's Return: Using a dedicated critic allows the model to compute token-level advantages, which is crucial when tracing complex, long-horizon decision trees.
• –Compaction Support: Group-wise methods struggle with comparing relative outputs when traces are compacted; the critic-based setup natively supports compacted sub-traces.
• –Agentic Specialization: This structural change positions GLM-5.2 as a highly capable model for project-scale engineering and multi-step agent actions.